Hierarchical Imitation Learning via Subgoal Representation Learning for Dynamic Treatment Recommendation